Hierarchical on-premises ai governance system and method

Abstract

A hierarchical ground-based artificial intelligence governance system and method are characterized by: user nodes actively pulling governance rule difference update packages from the central governance node and performing digital signature verification; the central governance node does not actively initiate connections to any user nodes; user nodes perform all data querying, document processing, and inference operations at the ground level, without transmitting them to the central governance node; user nodes perform query authorization determination, vector retrieval, ground-based artificial intelligence model inference, sensitive information detection, and multi-level output control according to ground-based governance rules; user nodes periodically upload anonymized trajectory summaries to the central governance node for governance status aggregation. This invention can solve the technical problems of data sovereignty protection, governance rule version consistency, update security verification, and compliance auditing in multi-user deployment environments.

Claims

1. A hierarchical ground-based artificial intelligence governance system, comprising a central governance node and at least one user node, wherein: The central governance node includes: a governance rule management module for establishing and versioning governance rule sets, each version of which has a unique version identifier; a model version management module for managing at least one AI model version, providing differential update packages and version description files; a rule distribution module for receiving pull requests from user nodes, comparing the local governance rule version number and AI model version number provided by the user node, and returning the corresponding differential update package and corresponding digital signature; and a governance status aggregation module for receiving an anonymized trajectory summary periodically uploaded by the user node, aggregating it to generate cross-node governance status data, and providing it to the user node the next time it pulls data; the user node includes: an active pull module for proactively initiating encrypted connection requests to the central governance node according to a set period or user-triggered conditions, transmitting the local governance rule version number and AI model version number to obtain the corresponding differential update package; An update verification module performs digital signature verification and version integrity check on the received difference update package. If the verification passes, the update is executed; if the verification fails, the update is automatically rejected and the anomaly is recorded. A query receiving module receives query requests from one of the authorized users and records the query initiation time and the anonymized user identifier. A rule determination module determines whether the query content is within the authorized processing scope according to the set of governance rules. If it does not meet the requirements, a rejection response is directly returned and the rejection trajectory is recorded. A vector retrieval module converts the authorized query into a vector representation, performs semantic similarity retrieval in one of the user node's vector databases, obtains a set of relevant file fragments, and executes the entire process on the user node's local end. A local inference module inputs the query and the set of file fragments into the user node's artificial intelligence model, performs inference operations, and generates a draft response content, which is executed entirely on the user node's local end. A sensitive information detection module detects the sensitive information type of the draft response content and performs hierarchical classification according to preset classification standards. An output control module performs detailed level restrictions and reference scope control on the response content according to the governance rule set, generating a final response content that conforms to the output specification; and a trajectory recording module records complete processing trajectory information and periodically generates the anonymized trajectory summary, which is actively uploaded to the central governance node; wherein, the user node actively initiates a connection request to the central governance node using the active pull module, and the central governance node does not actively initiate any network connection to the user node throughout the process; the user node's query data, document data, and inference calculations are all performed only at the user node's local end and are not transmitted to the central governance node; and the user node's update verification module refuses to update when the digital signature verification fails.

2. A hierarchical ground-end artificial intelligence governance system as described in Requisition 1, where the central governance node does not perform persistent storage, semantic parsing, or business processing of the original query content or document data on the received connection packets, and discards the packet content immediately after completion of the version comparison.

3. A hierarchical ground-end AI governance system as described in Request 1, wherein the update verification module further contains a version one rollback mechanism that enables the user node to automatically or manually roll back to the set of governance rules and the previous valid version of the AI ​​model when a new version of the governance rule causes system anomalies.

4. A hierarchical ground-end artificial intelligence governance system as described in Requisition 1, where the difference update packet returned by the rule distribution module contains only the change content compared to the previous version, and not the complete set of rules.

5. A hierarchical ground-end artificial intelligence governance system as described in Request 1, wherein the sensitive information detection module classifies the response content according to at least two levels of sensitivity, the output control module performs varying degrees of content restriction or masking processing according to the corresponding levels.

6. A hierarchical ground-end artificial intelligence governance system as described in Requisition 1, wherein the connection request of the active pull module contains an authentication token, the difference update packet is provided by the central governance node after verifying the authentication token.

7. A hierarchical ground-end artificial intelligence governance system as described in Requisition 1, in which each user node independently maintains the vector database and the AI ​​model at the corresponding ground end, the query data, document data and inferential operations between different user nodes are completely isolated from each other and are not visible from each other.

8. A hierarchical ground-end AI governance system as described in Requisition 1, wherein the same user node may be run in parallel by a complex number of computing units in a cluster, sharing the set of governance rules, the vector database and the AI ​​model of the same user node to achieve high availability or load balancing; 9. A hierarchical ground-end artificial intelligence governance system as described in Request 1, wherein the active pull module is further used to initiate an information request to the central governance node, the content of which information request contains the governance rule update request, the artificial intelligence model update request, or is governed through the center The central governance node sends the corresponding packet to the user node only according to the content of the information request and does not involve analysis, interpretation, inference or processing of the internal information owned by the user node or the artificial intelligence model.

10. A hierarchical ground-based artificial intelligence governance method, comprising the following steps: A user node, using an active pull module, initiates an encrypted connection request with authentication to a central governance node according to set conditions; The user node transmits a ground-based governance rule version number and an artificial intelligence model version number to the central governance node; The central governance node uses a rule distribution module to compare the differences between the ground-based governance rule version number and the artificial intelligence model version number. If both the ground-based governance rule version number and the artificial intelligence model version number are the latest versions, a no-update notification is sent back; otherwise, a difference update package and a corresponding digital signature are sent back; The user node uses an update verification module to perform digital signature verification and version integrity verification. If the verification passes, the user node receives the difference update package and updates the ground-based governance rule and the artificial intelligence model, and records the version change trajectory; if the verification fails, the user node refuses to update and records the abnormal trajectory; and the user node receives a query request from a user, determines the authorization scope of the query request according to the ground-based governance rule set, and directly sends a rejection response if it does not meet the requirements. For those who match the criteria, vector retrieval and AI model inference are performed at the user node, generating a draft response. Sensitive information detection and output control are then applied to this draft response to generate a final response. The processing trajectory is recorded, and an anonymized summary of the trajectory (excluding the original data) is periodically uploaded to the central governance node. The original query content and document data of that user node are not transmitted to the central governance node throughout the whole process; 11. The hierarchical ground-based artificial intelligence governance method as described in claim 10, wherein the digital signature verification further includes performing a version integrity check on the differential update package to confirm that the version of the differential update package has not been tampered with.

12. The hierarchical ground-based artificial intelligence governance method as described in claim 10, wherein the anonymized trajectory summary includes statistical information of query processing results, trigger frequency of sensitive information detection, and hierarchical distribution of output control execution, but does not contain any information that can restore the original query content.

Images